What is a large language model (LLM)?
A large language model (LLM) is a neural network trained on enormous amounts of text to predict what comes next in a sequence. That single skill, scaled up, turns out to cover writing, answering questions, summarizing, translating, and generating working code. LLMs are the engine behind ChatGPT, Claude, and Gemini, and behind every vibe coding tool.
"Large" refers to scale on both sides: billions of internal parameters, trained on datasets spanning much of the public internet. Scale is what separates an LLM from earlier language software; the same architecture at small size can't write a working app.
How does an LLM work?
An LLM reads and writes in tokens, small chunks of text roughly three-quarters of a word long in English. Given a sequence of tokens, the model predicts the most likely next one, appends it, and repeats. The architecture behind modern LLMs, the transformer, was introduced by Google researchers in the 2017 paper "Attention Is All You Need."
The output is probabilistic rather than fixed: the model samples from likely continuations, so the same prompt can produce different responses on different runs. What the model can consider at once is bounded by its context window, the working memory that holds the conversation, instructions, and files.
How does an LLM differ from traditional software?
Traditional software follows rules a programmer wrote: the same input produces the same output, and behavior can be traced to a line of code. An LLM's behavior emerges from training rather than explicit rules. There's no line of code that decides an answer, no lookup table of facts, and no guarantee of repeatability.
That difference changes how the technology is used responsibly. Rule-based software is verified by reading the rules; LLM output is verified by checking the result. For builders, that means treating each generated answer as a draft from a fast, capable, occasionally wrong collaborator, not as the deterministic output of a calculator.
What can LLMs do in software building?
LLMs generate code from natural-language descriptions, explain unfamiliar code in plain terms, diagnose errors from pasted stack traces, and translate logic between languages and frameworks. Vibe coding tools wrap an LLM in an agent: a loop that lets the model edit files, run commands, and read the results, turning text prediction into working software.
The model is also a capable reviewer of its own output when asked. Requests like "audit this code for exposed secrets" or "list the edge cases this form doesn't handle" reuse the same capability that generated the code to inspect it.
What are the limits of LLMs?
LLMs hallucinate: they produce confident, plausible output that's simply wrong, from invented functions to fabricated facts. They don't verify their own claims, their knowledge stops at a training cutoff date, and their context window caps how much they can hold at once. None of these limits announce themselves in the output.
That's why verification stays a human job. An LLM can write the code, and it can help check the code, but deciding whether the result is actually correct, secure, and fit for real users remains outside the model.
How do LLMs relate to AI agents?
The LLM is the model; an agent is the model plus tools plus a loop. An agent takes a goal, plans steps, calls tools (editing files, running tests, browsing), reads the results, and continues until done. Claude Code, Cursor, and similar tools are agents built on LLMs, which is why the same underlying model feels far more capable inside them than in a plain chat window.
